Robust algorithms of the type of stochastic approximation (continuous time)

The paper considers the problem of estimating an unknown drift parameter $\theta$ with observations $yt=\theta+\xi_t$ where $\xi_t$ is a stationary ergodic process. We prove strong consistency and asymptotic normality for the nonlinear estimation of the type of stochastic approximation
$$ \hat\theta=\theta_0+\int_0^t\frac{H(y_s-\hat\theta_s)}{(1+s)a_s}\,ds. $$
A method of choosing optimal (in the sense of limit variance) estimation of a function $H$ is offered.